Violence rates are higher in the South.
Historically, criminologists have turned to the idea of a Southern subculture of violence to explain this disparity.
But culture isn’t the only factor that varies regionally: the climate does, too.
Figure 1. Average yearly homicide rate per 100,000 by county, 1989–1991 (data from Anselin 2003) with Empirical Bayes smoothing; box map (hinge = 1.5IQR).
Climate varies by region.
This map of distinct climate regions (Ochoa 2018) is constructed from thirty years of climate data (Arguez et al. 2010).
The National Neighborhood Crime Study (Peterson & Krivo 2010) contains data for all Census tracts in a sample of cities.
The tracts with data for violent crime rate (averaged over 1999-2001) are located in fourteen climate regions, as shown in Figure 2.
Figure 2. Climate regions where sampled tracts are located.
Violence rates vary by climate region.
Figure 3 shows the climate-region coefficient, relative to region 1 (Great Lakes, in white), for the violent crime rate of the sampled tracts in each region.
Note that tracts in some Southern regions (like 6, 11, and 18) have higher violence rates on average than tracts in the Great Lakes region, but tracts in other Southern regions (like 8, 10, and 14) have lower violence rates.
This shows that the South is not homogeneous: there is variability across climate regions within the South.
Also note that tracts in region 3 have higher violence rates on average than those in the Great Lakes region and in the Florida Peninsula region (6).
Figure 3. Climate-region coefficient, relative to region 1 (Great Lakes), for violent crime rate of the sampled tracts in each region.
Controlling for theoretically and empirically indicated covariates changes the association between climate and violence.
Figure 4 shows the climate-region coefficient, relative to region 1 (Great Lakes), for violent crime rate of the sampled tracts in each region after controlling for tract- and city-level covariates.
This shows that once such conditions are taken into account, there are tracts in areas of the South that have low levels of violence (like in regions 8, 10, and 14) and tracts in areas outside the South that have high levels of violence (like in regions 5 and 12).
Figure 4. Model A: Climate-region coefficient, relative to region 1 (Great Lakes), for violent crime rate of the sampled tracts in each region.
In the absence of disadvantage, tracts in some subregions of the South exhibit high levels of violence—but tracts in other subregions have lower violence rates.
Controlling for the interaction of climate region and disadvantage allows the measurement of the association between climate region and violence when disadvantage is equal to zero.
Figure 5 shows that climate is related to high levels of violence for tracts in zone 18, the Lower South, but also for tracts in zone 5, the Pacific Northwest.
Furthermore, tracts in several Southern subregions have much lower violence rates under such conditions.
Together, these findings call into question the existence of a Southern subculture of violence.
Figure 5. Model B: Climate-region coefficient (when disadvantage = 0), relative to region 1, for violent crime rate of the sampled tracts in each region.
Climate moderates the relationship between disadvantage and violence.
Figure 6 shows that the regional climate changes the relationship between disadvantage and violent crime rate.
For tracts in regions 2, 3, 5, and 6, disadvantage is associated with high levels of violence; for regions 7, 10, 13, 14, 18, disadvantage is associated with lower levels of violence compared to region 1.
The results indicate that no matter the climate, disadvantage is positively related to violent crime, though the size of the effect varies by climate region.
Figure 6. Model B: Coefficient for the interaction between climate region and disadvantage, relative to region 1, for violent crime rate of the sampled tracts in each region.
Under a non-spatial regimes model, the baseline level of violence varies by climate region.
Model C controls for tract-level covariates, conducting fourteen region-specific OLS regressions. The model is correctly specified for three regions (8, 14, and 20); in the other regions, spatial lag, spatial error, or both are indicated.
Figure 7 shows that when all the covariates are equal to zero, the baseline level of violent crime varies by region: the Southern Plains zone (20) has a much higher baseline violence rate than the Texas Coast (8) or the Gulf Coast (14).
Figure 7. Model C, OLS by climate regions: Intercept, violent crime rate of the sampled tracts in each region.
Under a non-spatial regimes model, disadvantage is not associated with high levels of violence for tracts in some Southern subregions.
Figure 8. Model C, OLS by climate regions: Disadvantage coefficient, violent crime rate of the sampled tracts in each region.
Under a spatial lag model, the baseline level of violence varies by climate region.
Model D controls for tract-level covariates and the spatial lag of tract-level violent crime rate.
Figure 9 shows that for regions where this model is the correct specification, the baseline level of violence varies by climate region: when all covariates are equal to zero, the violent crime level of tracts in the Lower South zone (18) is much higher than that of tracts in the other regions.
Interestingly, the baseline violence rate for tracts in the Pacific Northwest (5) is higher than that of tracts in the Central Snake (2) and Midwest (3).
Figure 9. Model D, spatial lag by climate regions: Intercept, violent crime rate of the sampled tracts in each region.
Under a spatial lag model, the association between disadvantage and violence is not confined to the South, though it is strongest in the Lower South.
Figure 10 shows that violence in tracts in the Lower South (18) is driven by disadvantage, but this is also true, though to a lesser extent, for tracts in regions 2, 3, and 5.
Figure 10. Model D, spatial lag by regions: Disadvantage coefficient, violent crime rate of the sampled tracts in each region.
Under a spatial error model, the baseline level of violence varies by climate region.
Model E is a spatial error by regimes model that controls for tract-level covariates. The error term is significant for all regions, indicating that there are important unidentified covariates missing from the model.
Figure 11 shows that though tracts in the Florida Peninsula (6) have the highest baseline level of violent crime, tracts in the Northern Plains (13) also have a high baseline violence level.
Figure 11. Model E, spatial error by region: Intercept, violent crime rate of the sampled tracts in each region.
Under a spatial error model, violent crime of tracts in the Florida Peninsula is driven by disadvantage.
The high violence rate of tracts located in the Florida Peninsula zone (6) is associated with disadvantage, while the lower yet still high rates of tracts in regions 7, 10, and 13 are not associated with disadvantage.
Conclusions
Figure 11. Model E, spatial error by regions: Disadvantage coefficient, violent crime rate of the sampled tracts in each region.